ISSN1006-7167CN31-1707/TRESEARCHANDEXPLORATIONINLABORATORY第42卷第2期Vol.42No.22023年2月Feb.2023·实验技术·DOI:10.19927/j.cnki.syyt.2023.02.001基于深度学习的机械臂抓取姿态估计实验设计辛菁,龚爱玲,赵永红,穆凌霞,弋英民,张晓晖(西安理工大学自动化与信息工程学院,西安710048)摘要:为高效准确估计出机械臂对未知物体抓取姿态,提出了一种基于深度学习的机械臂抓取姿态估计方法,并设计了相关实验。该方法将生成抓取卷积网络GGCNN与挤压激励(SE)网络模块相结合,提出基于注意力机制的生成抓取卷积网络SE-GGCNN模型,提高抓取姿态估计的准确率。在Cornell数据集和JACQUARD数据集上进行了比较性实验。结果表明,相比于基本的GGCNN模型,SE-GGCNN模型在保证实时性的同时,将IoU指标值由原先的76%提升至82%;对于数据集中未出现过的新对象具备很好的鲁棒性和自适应性;单张图片115ms的检测用时表明所提出的方法适合实时应用,提升了未知物体抓取姿态估计的准确率。关键词:机械臂;深度学习;最优抓取姿态估计;注意力机制中图分类号:TP241.2文献标志码:A文章编号:1006-7167(2023)02-0001-04ExperimentalDesignofGraspingPoseEstimationofManipulatorBasedonDeepLearningXINJing,GONGAiling,ZHAOYonghong,MULingxia,YIYingmin,ZHANGXiaohui(SchoolofAutomationandInformationEngineering,Xi’anUniversityofTechnology,Xi’an710048,China)Abstract:Inthepaper,amethodofgraspingposeestimationofmanipulatorbasedondeeplearningisproposed,andrelatedexperimentsaredesigned.Theproposedmethodcombinesthegenerativegraspingconvolutionalneuralnetworkwithsqueeze-and-excitationnetworks,andproposestheimprovedSE-GGCNNnetworkmodelbasedonattentionmechanism.Themodelcanimprovetheaccuracyofgraspingposeestimation.ThecomparativeexperimentalresultsonCornelldatasetandJACQUARDdatasetshowthatcomparedwiththebasicGGCNNmodel,theproposedSE-GGCNNmodelincreasestheintersectionoverunion(IoU)indexvaluefromtheoriginal76%to82%whileensuringthereal-timeperformance.TheSE-GGCNNmodelhasgoodrobustnessandadaptabilityfornewobjectsthathavenotappearedinthedataset.Thedetectiontimeis115msforasingleimagewhichshowsthattheproposedmethodissuitableforreal-timeapplications,andfurth...